Model profile
GPT Oss 20B
Evidence summary
GPT Oss 20B has an estimated overall rank of #81; its 90% source-sensitivity interval is #39–#180. Its behavior-only rank is #88; company governance moves the combined estimate to #81. Published evidence spans 16 evals and 7 of 7 behavior components. Its strongest relative result is HELM Safety (simple_safety_tests, #1 of 80); its weakest is Enkrypt AI Safety Leaderboard (toxicity_attack_non_success_rate, #239 of 258).
Compare this model
Only models sharing at least one published sub-eval are listed.
Official and reference links
- Artificial Analysis ↗gpt-oss-20b
- Hugging Face ↗openai/gpt-oss-20b
- NVIDIA ↗qc69jvmznzxy/gpt-oss-20b
- OpenRouter ↗openai/gpt-oss-20b
- Model card ↗Family-level model document · OpenAI · first party
- Release source ↗direct research preview identity date
Published eval results
Rank is within that sub-eval. Black marks the observed result; the grey dot marks the value implied by the global rank. Values stay on each source’s native scale.
| Eval / sub-eval | Rank | Value | Distribution | Better | Source |
|---|---|---|---|---|---|
| AA-Omnisciencehallucination_rate | #212 / 311 | 0.866 | ↓ lower | Source ↗official | |
| AIRBench 2024 Safety Scenariossafety_scenarios | #10 / 80 | 0.86 | ↑ higher | Source ↗official | |
| Cisco AI Defense Rolling Single-Turn Leaderboardsingle_turn_attack_success_rate | #30 / 105 | 14.17 | ↓ lower | Source ↗official | |
| DystopiaBenchbasaglia_score | #42 / 50 | 70.13 | ↓ lower | Source ↗official | |
| DystopiaBenchbaudrillard_score | #37 / 50 | 66.3 | ↓ lower | Source ↗official | |
| DystopiaBenchhuxley_score | #37 / 50 | 75.47 | ↓ lower | Source ↗official | |
| DystopiaBenchlaguardia_score | #31 / 50 | 68.27 | ↓ lower | Source ↗official | |
| DystopiaBenchorwell_score | #35 / 50 | 73.53 | ↓ lower | Source ↗official | |
| DystopiaBenchpetrov_score | #28 / 50 | 73.53 | ↓ lower | Source ↗official | |
| Enkrypt AI Safety Leaderboardbias_attack_non_success_rate | #54 / 260 | 27.65 | ↑ higher | Source ↗official | |
| Enkrypt AI Safety Leaderboardcbrn_attack_non_success_rate | #213 / 260 | 81 | ↑ higher | Source ↗official | |
| Enkrypt AI Safety Leaderboardharmful_attack_non_success_rate | #94 / 260 | 80.56 | ↑ higher | Source ↗official | |
| Enkrypt AI Safety Leaderboardtoxicity_attack_non_success_rate | #239 / 258 | 80 | ↑ higher | Source ↗official | |
| FORTRESSaverage_risk_score | #15 / 49 | 17.62 | ↓ lower | Source ↗official | |
| FORTRESSover_refusal_score | #41 / 46 | 11.94 | ↓ lower | Source ↗official | |
| HELM Safetyanthropic_red_team | #10 / 80 | 0.997 | ↑ higher | Source ↗official | |
| HELM Safetybbq | #13 / 80 | 0.967 | ↑ higher | Source ↗official | |
| HELM Safetyharmbench | #2 / 80 | 0.987 | ↑ higher | Source ↗official | |
| HELM Safetysimple_safety_tests | #1 / 80 | 1 | ↑ higher | Source ↗official | |
| HELM Safetyxstest | #72 / 80 | 0.872 | ↑ higher | Source ↗official | |
| Manager Coercion Benchcoercion_ladder_depth | #11 / 31 | 7.9 | ↓ lower | Source ↗self run | |
| MASKlying_probability_pct | #10 / 53 | 13.54 | ↓ lower | Source ↗official | |
| ODCV-Benchaverage_severity | #7 / 12 | 1.475 | ↓ lower | Source ↗official | |
| ODCV-Benchmisalignment_rate | #5 / 12 | 28.75 | ↓ lower | Source ↗official | |
| SpeciEvalbelief_animal_sentience | #58 / 102 | 6.75 | ↑ higher | Source ↗official | |
| SpeciEvalland_animal_4ns | #12 / 102 | 4.2 | ↓ lower | Source ↗official | |
| SpeciEvalsea_animal_4ns | #59 / 102 | 4.83 | ↓ lower | Source ↗official | |
| SpeciEvalspeciesism | #40 / 102 | 1.9 | ↓ lower | Source ↗official | |
| Vigil Mental Health Safetyoverall_score | #17 / 23 | 31 | ↑ higher | Source ↗official |
Values evaluations
Descriptive values and political-framing results are separate from safety/ethics ranks. Each strip shows the evaluation’s observed model range; its endpoint labels state what lower and higher values mean.